[mpc00620]

Fixes #17.

TaskInitializer.init() loads the model and experiment configurations and schedules the feature registry setup and the periodic model-result and bloom-filter loading. Before this change, callers had no way to tell when the model data was in place, and the README told them to Thread.sleep(10000), which does not cover the first load: that load runs after a random delay of up to one refresh period (300 s by default).

This PR keeps that first load lazy, exactly as on main, and adds a readiness signal instead.

Changes:

• BidRequestEvaluatorFactory.isReady(). It is true once init() has completed, the extractor and transformer registries hold every ServiceLoader provider, and every model-result and bloom-filter loading task of that TaskInitializer has finished its first periodic run. The outcome of that run does not matter: a missing file, an S3 error or a loader exception still counts, so model data can never block readiness forever. Once true it stays true. The base class method throws UnsupportedOperationException, so existing subclasses keep compiling. DefaultBidRequestEvaluatorFactory overrides it.
• PeriodicTaskWithRandomizedStart.hasFinishedFirstRun(). This is a volatile flag set in a finally block of the scheduled run, so an exception or an Error still marks the run as finished.
• A TaskInitializer constructor that takes a callback. init() runs the callback once both stages have completed. init() now also throws when a stage ends with tasks still pending, instead of returning normally. The three-argument constructor is unchanged.
• getTask() on InitializerTaskOnPeriodicTask and ModelResultPeriodicLoadingInitializerTask.
• Javadoc for init(), isReady() and the loading tasks now says what actually happens. The README replaces the sleep with isReady() and says it can take up to one refresh period plus the load time.
• The docs also say what an unloaded model returns. It returns its no-match value: 1.0 (forward) for a LowValue model and 0.0 (low value) for a HighValue model. So evaluations fail open only for LowValue models. This is unchanged from main and is only documented here.

Tests:

• DefaultBidRequestEvaluatorFactoryReadinessTest uses a fake S3 and an executor that captures periodic tasks so each test decides when they run. It covers:
  • isReady() is false after init() and true only after both model tasks have run.
  • init() itself does not load model data.
  • A missing model file, or a loader failure, still makes isReady() true.
  • A configuration with no bloom-filter model is not blocked by the bloom-filter task.
  • isReady() stays true after a later refresh fails.
  • isReady() waits for the registry setup.
  • isReady() stays false when init() fails.
• PeriodicTaskWithRandomizedStartTest: the flag before and after the first run, after an Exception and after an Error, and the first delay staying within one period.
• TaskInitializerTest: the callback runs after both stages, is not run on failure, and a null callback is rejected.

LC_ALL=C.UTF-8 ./gradlew clean build passes: 556 tests, checkstyle and spotbugs green.